Unified deep potential model for hydrogen adsorption on pristine and Ti-decorated graphene

Abstract

Hydrogen interacts with pristine graphene mainly through weak dispersion forces, whereas Ti decoration considerably strengthens adsorption. These two systems therefore provide a convenient test for evaluating whether a single machine-learning interatomic potential can describe qualitatively different adsorption regimes without changing its parameterization.
To examine this problem, Deep Potential models were trained using density functional theory calculations with dispersion corrections for hydrogen adsorption on pristine and Ti-decorated graphene. Both descriptor-based and attention-based architectures were considered and trained using the same active-learning dataset.
The models were assessed by comparing adsorption energy profiles obtained from Deep Potential and density functional theory calculations. Particular attention was paid to the ability of the models to reproduce the weak physisorption potential of pristine graphene together with the stronger interaction on the Ti-decorated surface. The attention-based models provided the best agreement with the reference calculations for both systems, indicating that a single Deep Potential parameterization can describe adsorption over a broad interaction-energy range.
These results demonstrate that attention-based Deep Potential models are sufficiently transferable to describe hydrogen adsorption on carbon materials with substantially different interaction strengths, making them suitable for large-scale atomistic simulations.
The study was supported by the Russian Science Foundation (grant no. 25-22-20023, https://rscf.ru/en/project/25-22-20023/).

Speaker

Sozykin Sergey Anatolevich
South Ural State University
Russian Federation

Discussion

Ask question